ModelCensusopen-source ai reliability harness
The finding · Class 4 · Reasoning & Calculation

Real context moved 6.0pp on Method/Execution Split.

Picks the right method, then does the arithmetic wrong.

Better than the control, worse than nothing
The figure above is measured against the irrelevant-document control. Against no context at all this card is worse: 0% bare against 6% with the best supplied source. What the source repaired is damage the presence of a document did.

~intervals overlap — not separated from no change

The model chooses a correct method but executes it wrongly (or vice versa) — a right-approach/wrong-answer or wrong-approach/right-answer split. The rule

meta-llama/llama-3.1-8b-instructfmi_4_15 repeatsfmi_4_1_v1
In one frame
screenshot this
Method/Execution Splitfmi_4_1
meta-llama/llama-3.1-8b-instruct
Same question, both times
A shop sells 17 boxes holding 24 items each, then 8 items are returned. How many items were sold?
No contextPASS

To find the number of items sold, we need to multiply the number of boxes sold by the number of items in each box, then subtract the number of items returned. Number of boxes sold = 17 Number of items in each box = 24 Total items sold = 17 x 24 = 408 Items returned = 8 Items sold = 408 - 8 = 400 Answer: 400

detector fmi_4_1_v1: answer matches oracle

Curated contextPASS

To find the total number of items sold, we first need to find the total number of items sold initially. Total items initially = Number of boxes * Number of items per box Total items initially = 17 * 24 Total items initially = 408 Then, we subtract the number of items returned from the total number of items initially sold. Items sold = Total items initially - Items returned Items sold = 408 - 8 Items sold = 400 Answer: 400

detector fmi_4_1_v1: answer matches oracle

Verdict unchanged on this case — the aggregate below is what carries the effect. Across 5 repeats: +6.0pp (intervals overlap).fmi_4_1_v1 · meta-llama/llama-3.1-8b-instruct · 2026-09-09 · ground truth as of 2026-08-09

loop 2026-09-09 · meta-llama/llama-3.1-8b-instruct as served · open weights · released 2024-07 · 8B · set s1.4-live · cases v1.7 · conditions v1.0 · ground truth as of 2026-08-09

From no context to real context
5 repeats per case
failure rate · left is fewer failures0%10%20%30%40%No context0%[0%10%] n=50 · eff 35Irrelevant context0%presence — a document being there[0%10%] n=50 · eff 350.0pp OVERLAPSCurated context6%content — the information itself[2%19%] n=50 · eff 35+6.0pp OVERLAPS
whiskers are 95% intervals · each step is measured from the stop before it, not from zero
Try the other contexts
stored · no model call
1Questionheld fixed
A shop sells 17 boxes holding 24 items each, then 8 items are returned. How many items were sold?

expected as of 2026-08-09

400. Order matters: multiply before subtracting.

2Contextchange me

Nothing supplied. The baseline.

3ResponseNo context
PASS

To find the number of items sold, we need to multiply the number of boxes sold by the number of items in each box, then subtract the number of items returned. Number of boxes sold = 17 Number of items in each box = 24 Total items sold = 17 x 24 = 408 Items returned = 8 Items sold = 408 - 8 = 400 Answer: 400

fmi_4_1_v1

detector fmi_4_1_v1: answer matches oracle

The scoring
Failure rate by context
No context0% [0%10%] n=50 · eff 35
Irrelevant context0% [0%10%] n=50 · eff 35
General context6% [2%19%] n=49 · eff 34
Curated context6% [2%19%] n=50 · eff 35

0–40% · whisker = 95% interval

Trap vs answerable
No contexttrap 0%answerable 0%
Irrelevant contexttrap 0%answerable 0%
General contexttrap 7%answerable 5%
Curated contexttrap 7%answerable 5%

never pooled · a trap ratio would skew the delta

Residual — what survives

Supplying context did not reduce this failure, it increased it: 0% with no context against 6% under the strongest supplied source. What survives is therefore not a remainder but a substitution — the failures under context are 3 probes the detector still rejected, dominated by: answer 63.83 ≠ oracle 66.5.

Mitigations from the index — the claim
  • tool-assisted computation
  • verify final against method

fmi_4_1_v1 · 5 repeats per case · best context curated context · rollups + residuals committed · probe log outside git · manifest hash ties them · method